A recent study conducted by the UK's AI Security Institute (AISI) highlights a significant issue in standard AI evaluations, which systematically underestimate the capabilities of AI agents. The research, covering seven benchmarks, indicates that limiting the compute budget can lead to misleading assessments. Notably, when the token budget for software engineering tasks was increased tenfold, success rates improved by approximately 25%. This finding suggests that newer AI models are particularly benefitting from expanded resources. AISI's analysis reveals that actual advancements in AI capabilities may be around 60% steeper than previously recorded, emphasizing the need for a reevaluation of current benchmarking practices. In this context, it is important to note that a comprehensive understanding of AI's potential requires a more nuanced approach to evaluation metrics.